Photovoltaics off the
beaten track

Potential of solar energy in highway turnoffs across Germany

Outline

Motivation

• Germany has for the longest time relied heavily on coal and gas (25% in 2010)

• Climates pledges require a rapid transition to renewable energy

• Wind parks often face opposition from local communities

• Solar energy is a promising alternative because of its low footprint

• Photovoltaics (PV) can be installed in places with little alternative use

  • Most prominently: Roofs
  • But also: Artifical lakes, parking lots, highways

Experimental PV installation spots

PV in highway turnoffs: The idea

• Highway turnoffs are unattractive for most uses
• The city of Tübingen realized that they could be utilized for PV
• After several years of legal hassle, a site at the B28 in Lustnau is now the city’s largest source of green energy

Scaling across Germany

• How much potential for such projects is yet unused?
• We seek to answer this question by taking a closer look at the state of Brandenburg

Satellite imagery of two highway turnoffs in Brandenburg

When is a turnoff suitable for PV?

• We held expert interviews with the Tübingen project lead and an expert geologist

https://solarbusinesshub.com/2015/04/10/bike-path-on-koreas-highway-covered-with-solar-panels/

Xinhua News Agency via Getty Images

Insights from pilot project “Lustnauer Ohren”

• Pilot project: 8 weeks construction and 8 years approval (DER SPIEGEL, 2022)
• Nowadays:
  • More awareness and pressure from the public
  • Direct communication with the public via social media when bureaucratic hurdles arise
  • Consequently, authorities urged not to insist on 50-year-old regulations

Derivation of the project’s objective

• Illustrate potentials for renewable energies (PV) in a user-friendly way
  • Online application as information desk for municipalities and citizens
  • Alternative to complex expert reports

Conclusions for contextual focus of our project

Driving factors for one-off costs due to construction and operating costs:

• Environmental assessment for compensation areas
  • Nature in a potential ear is under protection (§ 44 Naturschutz Gesetz)
  • PV plant is an encroachment on the landscape
  • Woody plants such as bushes and trees are assessed differently (i.a. according to size, age)
  • Compensation areas must be equivalent or greater and geographically close
  • If landowner of the ear has no free compensation areas, they need to be must be rented (lease fees)
    • landowner for federal roads is the Regional Council of the county (in German: Regierungsbezirk)
    • for motorways the Autobahn GmbH/ indirectly the federal government)

Note: The environmental assessment determines one-off costs and long-term profitability of the PV.

Conclusions for our project from insights (pilot project)

Driving factors for one-off costs due to construction and operating costs:

• Distance to grid
• Meteorological details: Sunshine hours
• Geographical details: Ground conditions and slope
• [Sky alignment of the road]
  • Due to the technical innovations of recent years, the factor is less important nowadays as PVs can be built more flexibly than in Lustnau (insight from interview with Prof. Thomas Scholten)

Economic Model (Felix)

How is score calculated?

Data (Felix)

Potentially cover the following: * Data acquisition

• Data preparation

• Data validation

• Data analysis

• Data visualization

Land cover classification

• Problem: precisely predict the areal distribution of relevant features
  • high vegetation, low vegetation, buildings, etc.
• Task: Image segmentation
  • Divide image into regions belonging to different semantic classes.
    • Task of pixel-based classification
    • Pixel-wise mask for each object in the image
  • Output (class per pixel) is the same size as the input (pixel)

Figure credit: Choi, U. (2021, March 19) Semantic Segmentation (FCN, U-Net, DeepLab V3+).

Training data - Postdam

Potsdam 2D Semantic Segmentation dataset for the German city Potsdam, provided by the International Society for Photogrammetry and Remote Sensing (ISPRS)

Label	Color
1.) Impervious surfaces	white
2.) Building	blue
3.) Low vegetation	light-blue
4.) Tree	green
5.) Car	yellow
6.) Clutter/background	red

Data Source: 2D Semantic Labeling Contest - Potsdam. Available online: https://www.isprs.org/education/benchmarks/UrbanSemLab/default.aspx (accessed on 14 December 2022).

Training data - Postdam

Advantages:

• Aerial images, not satellite images
  • High resolution (6,000 px × 6,000 px with 5 cm ground res.)
• Precise and reliable labeling
  • Best dataset for German pixel-based labels
  • Classes are useful for our segmentation problem

Disadvantages:

• Urban structures, whereas in ears more rural structures
  • Rural structures are characterised by greater contiguous areas, like forests
  • Narrow, detailed urban structure in training creates fuzzy predictions for ears
• Giving greater weight to classes “tree” and “low vegetation”
  • Improves detection of organic structures
  • Good in detecting single trees, but not forests

Training data - Postdam

Color Labels: White - Impervious surfaces; Light-blue - Low vegetation; Green - Tree

Training data - LoveDA (regions of Nanjing, Changzhou and Wuhan / China)

Land-cOVEr Domain Adaptive semantic segmentation (LoveDA) dataset + High spatial resolution land-cover mapping (spatial resolution of 30 cm) + Source Google Earth platform

Label	Color
1.) Background	white
2.) Building	blue
3.) Road	light-blue
4.) Water	green
5.) Barren	yellow
6.) Forest	red
7.) Agriculture	red
0.) Unknown	red

Training data - LoveDA (regions of Nanjing, Changzhou and Wuhan / China)

Advantages * Dataset differentiates between rural and urban images + More rural scenes, but quite balanced dataset (2713 urban, 3274 rural) * Dataset created with the intention of achieving model transferability + “[U]rban and rural scenes can show completely different geographical landscapes, and the inadequate [generalization] of these algorithms hinders city-level or national-level mapping” (Wang et al., 2022, p. 1) + Dataset encounters for: 1) multi-scale objects; 2) complex background samples; 3) inconsistent class distributions

Disadvantages * Lower resolution than images from Brandenburg (30 cm vs. 20 cm, respectively) * Satellite images (vs. aerial images from Brandenburg) * [Less detailed labeling compared to Potsdam dataset] + Useful in training for the recognition of larger contiguous areas, but not individual trees

Semantic Segmentation - UNet

• Output (class per pixel) is the same size as the input (pixel)
• Network consists of a contracting path (downsampling) and an expansive path (upsampling) (Stevens et al., 2020b)
  • Downsampling:
    • Specific, detailed detectors (e.g. texture, color, characterising shapes)
    • Information reduction to relevant feature information
  • Upsampling:
  • Combination of the feature and spatial information
  • Extracting high-resolution features
• U-shaped network architecture: Early breakthrough for image segmentation

Model	Advantages	Disadv.
UNet	- widely used basic segmentation model	- training from scratch
	- feedback from interim presentation	- model overfits to training data
	- good starting point	- labelled data not specific for our task

Semantic Segmentation - ResNet50 + BigEarthNet (Sumbul et al., 2019)

• Residual networks (ResNets) allow very deep networks to be successfully trained (Stevens et al., 2020a)
  • Deep networks represent very complex functions
  • Problem with large networks: Gradient signal that goes to zero
  • Large number of multiplications, the gradient can decrease exponentially quickly to zero
  • Skip Connections: Additional layers do not hurt the model’s performance as regularisation will skip over them if those layers were not useful

Semantic Segmentation - ResNet50 + BigEarthNet (Sumbul et al., 2019)

• Transfer learning using pre-trained model backbones
  • Transfer knowledge learned from source dataset (BigEarthNet) to target dataset (training)
  • Model can already extract more general image features (e.g. textures, shapes)
• Pre-trained on BigEarthNet dataset
  • Atmospherically corrected satellite images from different seasons
  • 10 European countries
  • multiple land-cover classes (tile classifcation task, not segmentation)

Model	Advantages	Disadv.
ResNet50	- pre-trained weights	- trained for classification task
+ BigEarthNet	- seasonal satellite images	- inferior predictions
	- big dataset on European countries

Semantic Segmentation - ResNet101 + COCO (Lin et al., 2014)

• Pre-trained on COCO (Common Objects in Context) dataset
  • Large-scale object detection and segmentation dataset
  • Recognition in context (background awareness)

Model	Advantages	Disadv.
ResNet101	- pre-trained weights	- COCO dataset
+ COCO	- segmentation task	- trained for object segmentation
	- superior predictions after few training iterations
	- least overfitting
	- complexity of model seems to fit task

Semantic Segmentation - ResNet101 + COCO - Results

• Recognition of small forest areas very good.
• Individual trees are partly not recognized.
• Shadow areas are not always clearly assigned.

Color Labels: White - Impervious surfaces; Red - Road; Green - Trees

Semantic Segmentation - ResNet101 + COCO - Results

• Agricultural areas are better recognized when they are planted.
• Similar results for single trees that are leafy or bare.

Color Labels: White - Impervious surfaces; Red - Road; Green - Trees; Yellow - Agricultur

Semantic Segmentation - ResNet101 + COCO - Results

• Road recognition is so good that you could use it to identify the shape of an ear.
• Even small buildings in shaded areas are detected.
• The model is trained to recognise organic structures very well (characteristic for landscape in ears), however, human-made structures (fields, buildings) are therefore less regular.

Color Labels: White - Impervious surfaces; Red - Road; Green - Trees; Pink - Building

Presentation of the result - Dashboard (Jan)

References

• DER SPIEGEL. (2022, November 23, 0:50). Die Klima-Streber: Tübingens Weg in die Energiewende | SPIEGEL TV [Video]. YouTube. https://www.youtube.com/watch?v=WywER5wRLEI.
• Lin, T. Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., … & Zitnick, C. L. (2014). Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13 (pp. 740-755). Springer International Publishing.
• Stevens, E., Antiga, L., & Viehmann, T. (2020a). Using convolutions to generalize (Chapter 8). In Deep learning with PyTorch. Manning Publications.
• Stevens, E., Antiga, L., & Viehmann, T. (2020b). Using segmentation to find suspected nodules (Chapter 13). In Deep learning with PyTorch. Manning Publications.
• Sumbul, G., Charfuelan, M., Demir, B., & Markl, V. (2019). Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium (pp. 5901-5904). IEEE.
• Wang, J., Zheng, Z., Ma, A., Lu, X., & Zhong, Y. (2022). LoveDA: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733v6 [cs.CV].